OpenAI on Thursday announced the release of GPT-5.3-Codex-Spark, a lightweight version of its agentic coding model aimed at faster inference and real-time collaboration. The new model is powered by dedicated hardware from Cerebras Systems, signaling closer integration between OpenAI’s software and physical infrastructure.
Codex Spark is built to support rapid iteration and day-to-day coding tasks, rather than long-running workloads handled by the full GPT-5.3 model. OpenAI said the tool is optimized for low latency and is intended to improve responsiveness in collaborative development environments. The model is currently available as a research preview for ChatGPT Pro users through the Codex app.
The release marks the first product milestone in OpenAI’s multi-year partnership with Cerebras, announced last month and valued at more than $10 billion. Spark runs on Cerebras’ Wafer Scale Engine 3, a third-generation chip containing roughly 4 trillion transistors.
OpenAI said the new model represents an early step toward a dual-mode Codex platform, combining real-time coding assistance with deeper reasoning workflows. Cerebras has gained visibility in the AI hardware market and recently raised $1 billion at a reported $23 billion valuation.
